Safety-Aware Reinforcement Learning for Lunar Rover Navigation Using Orbital DEMs

Autonomous lunar rover navigation is constrained by Earth–Moon communication delay, sparse rewards, and high-variance polar terrain. This paper presents a DEM-driven digital twin framework for training and validating safety-aware rover navigation policies using NASA Lunar Reconnaissance Orbiter altimetry. The proposed system combines PPO with Potential-Based Reward Shaping, an onboard Attention U-Net hazard-perception module, a reactive safety filter, and a Multi-Axis Morphology-Aware Curriculum (MAMAC) that increases crater density, obstacle radius, corridor width, elevation variance, and path complexity during training. Across 20,410 in-distribution evaluation episodes, the integrated policy achieved 96% success on moderate terrain and 85% on high-complexity terrain, compared with 15% for heuristic-only baselines; collision rate on the hardest terrain decreased from 35% to 10%. We further evaluated 675 held-out DEM patches unseen during training: MAMAC reached 84.7% success at moderate difficulty and 53.0% at hard difficulty, versus 52.2% and 32.5% for a no-curriculum baseline. These results show that a curriculum-trained, perception-guided safety architecture provides an effective simulation-stage validation framework for autonomous lunar navigation.